This is a PyTorch re-implementation of MapNet, presented in:
João F. Henriques and Andrea Vedaldi, "MapNet: An Allocentric Spatial Memory for Mapping Environments", CVPR 2018 (PDF)
It reproduces all of the training from scratch for the mazes experiments, but not the Doom or AVD experiments; I hope to change that in the future.
Although it may work with older versions, this has mainly been tested with:
- PyTorch 1.3
- Python 3.7
- OverBoard 0.1.4 (for plotting and visualization)
The mazes are stored in a large text file (45 MB). For this reason, it is zipped in data/maze/mazes-10-10-100000.zip (6 MB), please extract its contents to the same directory.
Training can then be performed by running train_mapnet.py. Run train_mapnet.py --help for command-line options and their explanation.
Plots and tensor visualizations (mostly heatmaps of the joint position-orientation probability, as well as the maps) from OverBoard:
